[Question] Spatial buffers?

[idantene]

We've noticed that when using odc-stac with e.g.

query = stac_client.search(collections=["sentinel-2-l2a"],
                           bbox=wgs_bbox, datetime=[start_time, end_time],
                           limit=20)
items = query.item_collection()
stack = stac_load(items, bands=bands,
                  resolution=resolution, crs=f"EPSG:{epsg}", groupby="solar_day",
                  bbox=wgs_bbox, chunks={"x": -1, "y": -1}, stac_cfg=STAC_CFG,
                  fail_on_error=fail_on_errors)

We end up getting a slightly larger bounding box then the original bounding box (after converting from WGS to the desired CRS).
I wonder if this is intentional, and/or if there are explicit buffers in how STAC works? Or is it simply due to inaccuracies in WGS84 vs local projection?

When we later stitch several geotiffs together, this is not a particular concern, since the pixels overlap anyway and their values are identical. But when we ingest this data to an indexed database, we end up with duplicate errors that need to be handled 🤔

[Kirill888]

@idantene

this is as expected. But we still don't have a proper docs page that explains pixel grids to which to link to, for now see #95.

See anchor parameter of stac_load.

[idantene]

Thanks @Kirill888, that does make sense and I expected something of the sorts.

I'm not sure I follow how to use the anchor argument though. Should I use the AnchorEnum and turn grid snapping off completely, or use the geobox argument to specify the desired extents?

What would be the difference between the two behaviours, given that our bbox is a multiple of resolution in the target crs?

[idantene]

Also, may I suggest pinning #95 and #54? :)

[Kirill888]

@idantene

I'm not sure I follow how to use the anchor argument though. Should I use the AnchorEnum and turn grid snapping off completely, or use the geobox argument to specify the desired extents?

Specifying geobox= parameter is the most direct and precise way to control output. If you know bounding box and resolution you want you can just supply geobox=GeoBox.from_bbox(bbox, resolution=gsd, tight=True).

See GeoBox.from_bbox

What would be the difference between the two behaviours, given that our bbox is a multiple of resolution in the target crs?

Imagine axis aligned grid with spacing equal to the ground sampling distance, this grid is "anchored" tox=0 and y=0 (grid includes those two lines). Grid lines are pixel edges. This is the grid we "snap to" by default, that way all the image data loaded at the same resolution will have exactly overlapping pixels so you can just "paste" pixels from one array to another. When supplying anchor=AnchorEnum.FLOATING or tight=True this snapping is turned off, but you might still get an extra pixel on the right bottom, depending on floating point precision and resolution, you can control for that with tol= parameter of GeoBox.from_bbox, but we don't bubble up all those settings to odc.stac.load.

The main difference with supplying anchor= vs geobox= is that with anchor= you can have automatic decisions about resolution and CRS be made by odc-stac, while with geobox= you have to set those in your own code.